1 //===- InstCombineCompares.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitICmp and visitFCmp functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/IntrinsicInst.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/MemoryBuiltins.h"
18 #include "llvm/Target/TargetData.h"
19 #include "llvm/Support/ConstantRange.h"
20 #include "llvm/Support/GetElementPtrTypeIterator.h"
21 #include "llvm/Support/PatternMatch.h"
23 using namespace PatternMatch;
25 static ConstantInt *getOne(Constant *C) {
26 return ConstantInt::get(cast<IntegerType>(C->getType()), 1);
29 /// AddOne - Add one to a ConstantInt
30 static Constant *AddOne(Constant *C) {
31 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
33 /// SubOne - Subtract one from a ConstantInt
34 static Constant *SubOne(Constant *C) {
35 return ConstantExpr::getSub(C, ConstantInt::get(C->getType(), 1));
38 static ConstantInt *ExtractElement(Constant *V, Constant *Idx) {
39 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
42 static bool HasAddOverflow(ConstantInt *Result,
43 ConstantInt *In1, ConstantInt *In2,
46 if (In2->getValue().isNegative())
47 return Result->getValue().sgt(In1->getValue());
49 return Result->getValue().slt(In1->getValue());
51 return Result->getValue().ult(In1->getValue());
54 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
55 /// overflowed for this type.
56 static bool AddWithOverflow(Constant *&Result, Constant *In1,
57 Constant *In2, bool IsSigned = false) {
58 Result = ConstantExpr::getAdd(In1, In2);
60 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
61 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
62 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
63 if (HasAddOverflow(ExtractElement(Result, Idx),
64 ExtractElement(In1, Idx),
65 ExtractElement(In2, Idx),
72 return HasAddOverflow(cast<ConstantInt>(Result),
73 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
77 static bool HasSubOverflow(ConstantInt *Result,
78 ConstantInt *In1, ConstantInt *In2,
81 if (In2->getValue().isNegative())
82 return Result->getValue().slt(In1->getValue());
84 return Result->getValue().sgt(In1->getValue());
86 return Result->getValue().ugt(In1->getValue());
89 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
90 /// overflowed for this type.
91 static bool SubWithOverflow(Constant *&Result, Constant *In1,
92 Constant *In2, bool IsSigned = false) {
93 Result = ConstantExpr::getSub(In1, In2);
95 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
96 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
97 Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i);
98 if (HasSubOverflow(ExtractElement(Result, Idx),
99 ExtractElement(In1, Idx),
100 ExtractElement(In2, Idx),
107 return HasSubOverflow(cast<ConstantInt>(Result),
108 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
112 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
113 /// comparison only checks the sign bit. If it only checks the sign bit, set
114 /// TrueIfSigned if the result of the comparison is true when the input value is
116 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
117 bool &TrueIfSigned) {
119 case ICmpInst::ICMP_SLT: // True if LHS s< 0
121 return RHS->isZero();
122 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
124 return RHS->isAllOnesValue();
125 case ICmpInst::ICMP_SGT: // True if LHS s> -1
126 TrueIfSigned = false;
127 return RHS->isAllOnesValue();
128 case ICmpInst::ICMP_UGT:
129 // True if LHS u> RHS and RHS == high-bit-mask - 1
131 return RHS->getValue() ==
132 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
133 case ICmpInst::ICMP_UGE:
134 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
136 return RHS->getValue().isSignBit();
142 // isHighOnes - Return true if the constant is of the form 1+0+.
143 // This is the same as lowones(~X).
144 static bool isHighOnes(const ConstantInt *CI) {
145 return (~CI->getValue() + 1).isPowerOf2();
148 /// ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
149 /// set of known zero and one bits, compute the maximum and minimum values that
150 /// could have the specified known zero and known one bits, returning them in
152 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
153 const APInt& KnownOne,
154 APInt& Min, APInt& Max) {
155 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
156 KnownZero.getBitWidth() == Min.getBitWidth() &&
157 KnownZero.getBitWidth() == Max.getBitWidth() &&
158 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
159 APInt UnknownBits = ~(KnownZero|KnownOne);
161 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
162 // bit if it is unknown.
164 Max = KnownOne|UnknownBits;
166 if (UnknownBits.isNegative()) { // Sign bit is unknown
167 Min.setBit(Min.getBitWidth()-1);
168 Max.clearBit(Max.getBitWidth()-1);
172 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
173 // a set of known zero and one bits, compute the maximum and minimum values that
174 // could have the specified known zero and known one bits, returning them in
176 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
177 const APInt &KnownOne,
178 APInt &Min, APInt &Max) {
179 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
180 KnownZero.getBitWidth() == Min.getBitWidth() &&
181 KnownZero.getBitWidth() == Max.getBitWidth() &&
182 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
183 APInt UnknownBits = ~(KnownZero|KnownOne);
185 // The minimum value is when the unknown bits are all zeros.
187 // The maximum value is when the unknown bits are all ones.
188 Max = KnownOne|UnknownBits;
193 /// FoldCmpLoadFromIndexedGlobal - Called we see this pattern:
194 /// cmp pred (load (gep GV, ...)), cmpcst
195 /// where GV is a global variable with a constant initializer. Try to simplify
196 /// this into some simple computation that does not need the load. For example
197 /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3".
199 /// If AndCst is non-null, then the loaded value is masked with that constant
200 /// before doing the comparison. This handles cases like "A[i]&4 == 0".
201 Instruction *InstCombiner::
202 FoldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, GlobalVariable *GV,
203 CmpInst &ICI, ConstantInt *AndCst) {
204 // We need TD information to know the pointer size unless this is inbounds.
205 if (!GEP->isInBounds() && TD == 0) return 0;
207 ConstantArray *Init = dyn_cast<ConstantArray>(GV->getInitializer());
208 if (Init == 0 || Init->getNumOperands() > 1024) return 0;
210 // There are many forms of this optimization we can handle, for now, just do
211 // the simple index into a single-dimensional array.
213 // Require: GEP GV, 0, i {{, constant indices}}
214 if (GEP->getNumOperands() < 3 ||
215 !isa<ConstantInt>(GEP->getOperand(1)) ||
216 !cast<ConstantInt>(GEP->getOperand(1))->isZero() ||
217 isa<Constant>(GEP->getOperand(2)))
220 // Check that indices after the variable are constants and in-range for the
221 // type they index. Collect the indices. This is typically for arrays of
223 SmallVector<unsigned, 4> LaterIndices;
225 const Type *EltTy = cast<ArrayType>(Init->getType())->getElementType();
226 for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) {
227 ConstantInt *Idx = dyn_cast<ConstantInt>(GEP->getOperand(i));
228 if (Idx == 0) return 0; // Variable index.
230 uint64_t IdxVal = Idx->getZExtValue();
231 if ((unsigned)IdxVal != IdxVal) return 0; // Too large array index.
233 if (const StructType *STy = dyn_cast<StructType>(EltTy))
234 EltTy = STy->getElementType(IdxVal);
235 else if (const ArrayType *ATy = dyn_cast<ArrayType>(EltTy)) {
236 if (IdxVal >= ATy->getNumElements()) return 0;
237 EltTy = ATy->getElementType();
239 return 0; // Unknown type.
242 LaterIndices.push_back(IdxVal);
245 enum { Overdefined = -3, Undefined = -2 };
247 // Variables for our state machines.
249 // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form
250 // "i == 47 | i == 87", where 47 is the first index the condition is true for,
251 // and 87 is the second (and last) index. FirstTrueElement is -2 when
252 // undefined, otherwise set to the first true element. SecondTrueElement is
253 // -2 when undefined, -3 when overdefined and >= 0 when that index is true.
254 int FirstTrueElement = Undefined, SecondTrueElement = Undefined;
256 // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the
257 // form "i != 47 & i != 87". Same state transitions as for true elements.
258 int FirstFalseElement = Undefined, SecondFalseElement = Undefined;
260 /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these
261 /// define a state machine that triggers for ranges of values that the index
262 /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'.
263 /// This is -2 when undefined, -3 when overdefined, and otherwise the last
264 /// index in the range (inclusive). We use -2 for undefined here because we
265 /// use relative comparisons and don't want 0-1 to match -1.
266 int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined;
268 // MagicBitvector - This is a magic bitvector where we set a bit if the
269 // comparison is true for element 'i'. If there are 64 elements or less in
270 // the array, this will fully represent all the comparison results.
271 uint64_t MagicBitvector = 0;
274 // Scan the array and see if one of our patterns matches.
275 Constant *CompareRHS = cast<Constant>(ICI.getOperand(1));
276 for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) {
277 Constant *Elt = Init->getOperand(i);
279 // If this is indexing an array of structures, get the structure element.
280 if (!LaterIndices.empty())
281 Elt = ConstantExpr::getExtractValue(Elt, LaterIndices.data(),
282 LaterIndices.size());
284 // If the element is masked, handle it.
285 if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst);
287 // Find out if the comparison would be true or false for the i'th element.
288 Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt,
290 // If the result is undef for this element, ignore it.
291 if (isa<UndefValue>(C)) {
292 // Extend range state machines to cover this element in case there is an
293 // undef in the middle of the range.
294 if (TrueRangeEnd == (int)i-1)
296 if (FalseRangeEnd == (int)i-1)
301 // If we can't compute the result for any of the elements, we have to give
302 // up evaluating the entire conditional.
303 if (!isa<ConstantInt>(C)) return 0;
305 // Otherwise, we know if the comparison is true or false for this element,
306 // update our state machines.
307 bool IsTrueForElt = !cast<ConstantInt>(C)->isZero();
309 // State machine for single/double/range index comparison.
311 // Update the TrueElement state machine.
312 if (FirstTrueElement == Undefined)
313 FirstTrueElement = TrueRangeEnd = i; // First true element.
315 // Update double-compare state machine.
316 if (SecondTrueElement == Undefined)
317 SecondTrueElement = i;
319 SecondTrueElement = Overdefined;
321 // Update range state machine.
322 if (TrueRangeEnd == (int)i-1)
325 TrueRangeEnd = Overdefined;
328 // Update the FalseElement state machine.
329 if (FirstFalseElement == Undefined)
330 FirstFalseElement = FalseRangeEnd = i; // First false element.
332 // Update double-compare state machine.
333 if (SecondFalseElement == Undefined)
334 SecondFalseElement = i;
336 SecondFalseElement = Overdefined;
338 // Update range state machine.
339 if (FalseRangeEnd == (int)i-1)
342 FalseRangeEnd = Overdefined;
347 // If this element is in range, update our magic bitvector.
348 if (i < 64 && IsTrueForElt)
349 MagicBitvector |= 1ULL << i;
351 // If all of our states become overdefined, bail out early. Since the
352 // predicate is expensive, only check it every 8 elements. This is only
353 // really useful for really huge arrays.
354 if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined &&
355 SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined &&
356 FalseRangeEnd == Overdefined)
360 // Now that we've scanned the entire array, emit our new comparison(s). We
361 // order the state machines in complexity of the generated code.
362 Value *Idx = GEP->getOperand(2);
364 // If the index is larger than the pointer size of the target, truncate the
365 // index down like the GEP would do implicitly. We don't have to do this for
366 // an inbounds GEP because the index can't be out of range.
367 if (!GEP->isInBounds() &&
368 Idx->getType()->getPrimitiveSizeInBits() > TD->getPointerSizeInBits())
369 Idx = Builder->CreateTrunc(Idx, TD->getIntPtrType(Idx->getContext()));
371 // If the comparison is only true for one or two elements, emit direct
373 if (SecondTrueElement != Overdefined) {
374 // None true -> false.
375 if (FirstTrueElement == Undefined)
376 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(GEP->getContext()));
378 Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement);
380 // True for one element -> 'i == 47'.
381 if (SecondTrueElement == Undefined)
382 return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx);
384 // True for two elements -> 'i == 47 | i == 72'.
385 Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx);
386 Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement);
387 Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx);
388 return BinaryOperator::CreateOr(C1, C2);
391 // If the comparison is only false for one or two elements, emit direct
393 if (SecondFalseElement != Overdefined) {
394 // None false -> true.
395 if (FirstFalseElement == Undefined)
396 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(GEP->getContext()));
398 Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement);
400 // False for one element -> 'i != 47'.
401 if (SecondFalseElement == Undefined)
402 return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx);
404 // False for two elements -> 'i != 47 & i != 72'.
405 Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx);
406 Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement);
407 Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx);
408 return BinaryOperator::CreateAnd(C1, C2);
411 // If the comparison can be replaced with a range comparison for the elements
412 // where it is true, emit the range check.
413 if (TrueRangeEnd != Overdefined) {
414 assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare");
416 // Generate (i-FirstTrue) <u (TrueRangeEnd-FirstTrue+1).
417 if (FirstTrueElement) {
418 Value *Offs = ConstantInt::get(Idx->getType(), -FirstTrueElement);
419 Idx = Builder->CreateAdd(Idx, Offs);
422 Value *End = ConstantInt::get(Idx->getType(),
423 TrueRangeEnd-FirstTrueElement+1);
424 return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End);
427 // False range check.
428 if (FalseRangeEnd != Overdefined) {
429 assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare");
430 // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse).
431 if (FirstFalseElement) {
432 Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement);
433 Idx = Builder->CreateAdd(Idx, Offs);
436 Value *End = ConstantInt::get(Idx->getType(),
437 FalseRangeEnd-FirstFalseElement);
438 return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End);
442 // If a 32-bit or 64-bit magic bitvector captures the entire comparison state
443 // of this load, replace it with computation that does:
444 // ((magic_cst >> i) & 1) != 0
445 if (Init->getNumOperands() <= 32 ||
446 (TD && Init->getNumOperands() <= 64 && TD->isLegalInteger(64))) {
448 if (Init->getNumOperands() <= 32)
449 Ty = Type::getInt32Ty(Init->getContext());
451 Ty = Type::getInt64Ty(Init->getContext());
452 Value *V = Builder->CreateIntCast(Idx, Ty, false);
453 V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V);
454 V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V);
455 return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0));
462 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
463 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
464 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
465 /// be complex, and scales are involved. The above expression would also be
466 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
467 /// This later form is less amenable to optimization though, and we are allowed
468 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
470 /// If we can't emit an optimized form for this expression, this returns null.
472 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
474 TargetData &TD = *IC.getTargetData();
475 gep_type_iterator GTI = gep_type_begin(GEP);
477 // Check to see if this gep only has a single variable index. If so, and if
478 // any constant indices are a multiple of its scale, then we can compute this
479 // in terms of the scale of the variable index. For example, if the GEP
480 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
481 // because the expression will cross zero at the same point.
482 unsigned i, e = GEP->getNumOperands();
484 for (i = 1; i != e; ++i, ++GTI) {
485 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
486 // Compute the aggregate offset of constant indices.
487 if (CI->isZero()) continue;
489 // Handle a struct index, which adds its field offset to the pointer.
490 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
491 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
493 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
494 Offset += Size*CI->getSExtValue();
497 // Found our variable index.
502 // If there are no variable indices, we must have a constant offset, just
503 // evaluate it the general way.
504 if (i == e) return 0;
506 Value *VariableIdx = GEP->getOperand(i);
507 // Determine the scale factor of the variable element. For example, this is
508 // 4 if the variable index is into an array of i32.
509 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
511 // Verify that there are no other variable indices. If so, emit the hard way.
512 for (++i, ++GTI; i != e; ++i, ++GTI) {
513 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
516 // Compute the aggregate offset of constant indices.
517 if (CI->isZero()) continue;
519 // Handle a struct index, which adds its field offset to the pointer.
520 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
521 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
523 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
524 Offset += Size*CI->getSExtValue();
528 // Okay, we know we have a single variable index, which must be a
529 // pointer/array/vector index. If there is no offset, life is simple, return
531 unsigned IntPtrWidth = TD.getPointerSizeInBits();
533 // Cast to intptrty in case a truncation occurs. If an extension is needed,
534 // we don't need to bother extending: the extension won't affect where the
535 // computation crosses zero.
536 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
537 VariableIdx = new TruncInst(VariableIdx,
538 TD.getIntPtrType(VariableIdx->getContext()),
539 VariableIdx->getName(), &I);
543 // Otherwise, there is an index. The computation we will do will be modulo
544 // the pointer size, so get it.
545 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
547 Offset &= PtrSizeMask;
548 VariableScale &= PtrSizeMask;
550 // To do this transformation, any constant index must be a multiple of the
551 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
552 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
553 // multiple of the variable scale.
554 int64_t NewOffs = Offset / (int64_t)VariableScale;
555 if (Offset != NewOffs*(int64_t)VariableScale)
558 // Okay, we can do this evaluation. Start by converting the index to intptr.
559 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
560 if (VariableIdx->getType() != IntPtrTy)
561 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
563 VariableIdx->getName(), &I);
564 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
565 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
568 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
569 /// else. At this point we know that the GEP is on the LHS of the comparison.
570 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
571 ICmpInst::Predicate Cond,
573 // Look through bitcasts.
574 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
575 RHS = BCI->getOperand(0);
577 Value *PtrBase = GEPLHS->getOperand(0);
578 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
579 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
580 // This transformation (ignoring the base and scales) is valid because we
581 // know pointers can't overflow since the gep is inbounds. See if we can
582 // output an optimized form.
583 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
585 // If not, synthesize the offset the hard way.
587 Offset = EmitGEPOffset(GEPLHS);
588 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
589 Constant::getNullValue(Offset->getType()));
590 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
591 // If the base pointers are different, but the indices are the same, just
592 // compare the base pointer.
593 if (PtrBase != GEPRHS->getOperand(0)) {
594 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
595 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
596 GEPRHS->getOperand(0)->getType();
598 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
599 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
600 IndicesTheSame = false;
604 // If all indices are the same, just compare the base pointers.
606 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
607 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
609 // Otherwise, the base pointers are different and the indices are
610 // different, bail out.
614 // If one of the GEPs has all zero indices, recurse.
615 bool AllZeros = true;
616 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
617 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
618 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
623 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
624 ICmpInst::getSwappedPredicate(Cond), I);
626 // If the other GEP has all zero indices, recurse.
628 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
629 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
630 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
635 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
637 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
638 // If the GEPs only differ by one index, compare it.
639 unsigned NumDifferences = 0; // Keep track of # differences.
640 unsigned DiffOperand = 0; // The operand that differs.
641 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
642 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
643 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
644 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
645 // Irreconcilable differences.
649 if (NumDifferences++) break;
654 if (NumDifferences == 0) // SAME GEP?
655 return ReplaceInstUsesWith(I, // No comparison is needed here.
656 ConstantInt::get(Type::getInt1Ty(I.getContext()),
657 ICmpInst::isTrueWhenEqual(Cond)));
659 else if (NumDifferences == 1) {
660 Value *LHSV = GEPLHS->getOperand(DiffOperand);
661 Value *RHSV = GEPRHS->getOperand(DiffOperand);
662 // Make sure we do a signed comparison here.
663 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
667 // Only lower this if the icmp is the only user of the GEP or if we expect
668 // the result to fold to a constant!
670 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
671 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
672 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
673 Value *L = EmitGEPOffset(GEPLHS);
674 Value *R = EmitGEPOffset(GEPRHS);
675 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
681 /// FoldICmpAddOpCst - Fold "icmp pred (X+CI), X".
682 Instruction *InstCombiner::FoldICmpAddOpCst(ICmpInst &ICI,
683 Value *X, ConstantInt *CI,
684 ICmpInst::Predicate Pred,
686 // If we have X+0, exit early (simplifying logic below) and let it get folded
687 // elsewhere. icmp X+0, X -> icmp X, X
689 bool isTrue = ICmpInst::isTrueWhenEqual(Pred);
690 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
693 // (X+4) == X -> false.
694 if (Pred == ICmpInst::ICMP_EQ)
695 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
697 // (X+4) != X -> true.
698 if (Pred == ICmpInst::ICMP_NE)
699 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
701 // If this is an instruction (as opposed to constantexpr) get NUW/NSW info.
702 bool isNUW = false, isNSW = false;
703 if (BinaryOperator *Add = dyn_cast<BinaryOperator>(TheAdd)) {
704 isNUW = Add->hasNoUnsignedWrap();
705 isNSW = Add->hasNoSignedWrap();
708 // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0,
709 // so the values can never be equal. Similiarly for all other "or equals"
712 // (X+1) <u X --> X >u (MAXUINT-1) --> X == 255
713 // (X+2) <u X --> X >u (MAXUINT-2) --> X > 253
714 // (X+MAXUINT) <u X --> X >u (MAXUINT-MAXUINT) --> X != 0
715 if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
716 // If this is an NUW add, then this is always false.
718 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(X->getContext()));
721 ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI);
722 return new ICmpInst(ICmpInst::ICMP_UGT, X, R);
725 // (X+1) >u X --> X <u (0-1) --> X != 255
726 // (X+2) >u X --> X <u (0-2) --> X <u 254
727 // (X+MAXUINT) >u X --> X <u (0-MAXUINT) --> X <u 1 --> X == 0
728 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
729 // If this is an NUW add, then this is always true.
731 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(X->getContext()));
732 return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI));
735 unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits();
736 ConstantInt *SMax = ConstantInt::get(X->getContext(),
737 APInt::getSignedMaxValue(BitWidth));
739 // (X+ 1) <s X --> X >s (MAXSINT-1) --> X == 127
740 // (X+ 2) <s X --> X >s (MAXSINT-2) --> X >s 125
741 // (X+MAXSINT) <s X --> X >s (MAXSINT-MAXSINT) --> X >s 0
742 // (X+MINSINT) <s X --> X >s (MAXSINT-MINSINT) --> X >s -1
743 // (X+ -2) <s X --> X >s (MAXSINT- -2) --> X >s 126
744 // (X+ -1) <s X --> X >s (MAXSINT- -1) --> X != 127
745 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
746 // If this is an NSW add, then we have two cases: if the constant is
747 // positive, then this is always false, if negative, this is always true.
749 bool isTrue = CI->getValue().isNegative();
750 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
753 return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI));
756 // (X+ 1) >s X --> X <s (MAXSINT-(1-1)) --> X != 127
757 // (X+ 2) >s X --> X <s (MAXSINT-(2-1)) --> X <s 126
758 // (X+MAXSINT) >s X --> X <s (MAXSINT-(MAXSINT-1)) --> X <s 1
759 // (X+MINSINT) >s X --> X <s (MAXSINT-(MINSINT-1)) --> X <s -2
760 // (X+ -2) >s X --> X <s (MAXSINT-(-2-1)) --> X <s -126
761 // (X+ -1) >s X --> X <s (MAXSINT-(-1-1)) --> X == -128
763 // If this is an NSW add, then we have two cases: if the constant is
764 // positive, then this is always true, if negative, this is always false.
766 bool isTrue = !CI->getValue().isNegative();
767 return ReplaceInstUsesWith(ICI, ConstantInt::get(ICI.getType(), isTrue));
770 assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE);
771 Constant *C = ConstantInt::get(X->getContext(), CI->getValue()-1);
772 return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C));
775 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
776 /// and CmpRHS are both known to be integer constants.
777 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
778 ConstantInt *DivRHS) {
779 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
780 const APInt &CmpRHSV = CmpRHS->getValue();
782 // FIXME: If the operand types don't match the type of the divide
783 // then don't attempt this transform. The code below doesn't have the
784 // logic to deal with a signed divide and an unsigned compare (and
785 // vice versa). This is because (x /s C1) <s C2 produces different
786 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
787 // (x /u C1) <u C2. Simply casting the operands and result won't
788 // work. :( The if statement below tests that condition and bails
790 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
791 if (!ICI.isEquality() && DivIsSigned != ICI.isSigned())
793 if (DivRHS->isZero())
794 return 0; // The ProdOV computation fails on divide by zero.
795 if (DivIsSigned && DivRHS->isAllOnesValue())
796 return 0; // The overflow computation also screws up here
798 return 0; // Not worth bothering, and eliminates some funny cases
801 // Compute Prod = CI * DivRHS. We are essentially solving an equation
802 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
803 // C2 (CI). By solving for X we can turn this into a range check
804 // instead of computing a divide.
805 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
807 // Determine if the product overflows by seeing if the product is
808 // not equal to the divide. Make sure we do the same kind of divide
809 // as in the LHS instruction that we're folding.
810 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
811 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
813 // Get the ICmp opcode
814 ICmpInst::Predicate Pred = ICI.getPredicate();
816 /// If the division is known to be exact, then there is no remainder from the
817 /// divide, so the covered range size is unit, otherwise it is the divisor.
818 ConstantInt *RangeSize = DivI->isExact() ? getOne(Prod) : DivRHS;
820 // Figure out the interval that is being checked. For example, a comparison
821 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
822 // Compute this interval based on the constants involved and the signedness of
823 // the compare/divide. This computes a half-open interval, keeping track of
824 // whether either value in the interval overflows. After analysis each
825 // overflow variable is set to 0 if it's corresponding bound variable is valid
826 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
827 int LoOverflow = 0, HiOverflow = 0;
828 Constant *LoBound = 0, *HiBound = 0;
830 if (!DivIsSigned) { // udiv
831 // e.g. X/5 op 3 --> [15, 20)
833 HiOverflow = LoOverflow = ProdOV;
835 // If this is not an exact divide, then many values in the range collapse
836 // to the same result value.
837 HiOverflow = AddWithOverflow(HiBound, LoBound, RangeSize, false);
840 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
841 if (CmpRHSV == 0) { // (X / pos) op 0
842 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
843 LoBound = ConstantExpr::getNeg(SubOne(RangeSize));
845 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
846 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
847 HiOverflow = LoOverflow = ProdOV;
849 HiOverflow = AddWithOverflow(HiBound, Prod, RangeSize, true);
850 } else { // (X / pos) op neg
851 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
852 HiBound = AddOne(Prod);
853 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
855 ConstantInt *DivNeg =cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
856 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0;
859 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
861 RangeSize = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
862 if (CmpRHSV == 0) { // (X / neg) op 0
863 // e.g. X/-5 op 0 --> [-4, 5)
864 LoBound = AddOne(RangeSize);
865 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(RangeSize));
866 if (HiBound == DivRHS) { // -INTMIN = INTMIN
867 HiOverflow = 1; // [INTMIN+1, overflow)
868 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
870 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
871 // e.g. X/-5 op 3 --> [-19, -14)
872 HiBound = AddOne(Prod);
873 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
875 LoOverflow = AddWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0;
876 } else { // (X / neg) op neg
877 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
878 LoOverflow = HiOverflow = ProdOV;
880 HiOverflow = SubWithOverflow(HiBound, Prod, RangeSize, true);
883 // Dividing by a negative swaps the condition. LT <-> GT
884 Pred = ICmpInst::getSwappedPredicate(Pred);
887 Value *X = DivI->getOperand(0);
889 default: llvm_unreachable("Unhandled icmp opcode!");
890 case ICmpInst::ICMP_EQ:
891 if (LoOverflow && HiOverflow)
892 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
894 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
895 ICmpInst::ICMP_UGE, X, LoBound);
897 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
898 ICmpInst::ICMP_ULT, X, HiBound);
899 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
901 case ICmpInst::ICMP_NE:
902 if (LoOverflow && HiOverflow)
903 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
905 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
906 ICmpInst::ICMP_ULT, X, LoBound);
908 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
909 ICmpInst::ICMP_UGE, X, HiBound);
910 return ReplaceInstUsesWith(ICI, InsertRangeTest(X, LoBound, HiBound,
911 DivIsSigned, false));
912 case ICmpInst::ICMP_ULT:
913 case ICmpInst::ICMP_SLT:
914 if (LoOverflow == +1) // Low bound is greater than input range.
915 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
916 if (LoOverflow == -1) // Low bound is less than input range.
917 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
918 return new ICmpInst(Pred, X, LoBound);
919 case ICmpInst::ICMP_UGT:
920 case ICmpInst::ICMP_SGT:
921 if (HiOverflow == +1) // High bound greater than input range.
922 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(ICI.getContext()));
923 if (HiOverflow == -1) // High bound less than input range.
924 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(ICI.getContext()));
925 if (Pred == ICmpInst::ICMP_UGT)
926 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
927 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
931 /// FoldICmpShrCst - Handle "icmp(([al]shr X, cst1), cst2)".
932 Instruction *InstCombiner::FoldICmpShrCst(ICmpInst &ICI, BinaryOperator *Shr,
933 ConstantInt *ShAmt) {
934 if (!ICI.isEquality()) return 0;
936 const APInt &CmpRHSV = cast<ConstantInt>(ICI.getOperand(1))->getValue();
938 // Check that the shift amount is in range. If not, don't perform
939 // undefined shifts. When the shift is visited it will be
941 uint32_t TypeBits = CmpRHSV.getBitWidth();
942 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
943 if (ShAmtVal >= TypeBits)
946 // If we are comparing against bits always shifted out, the
947 // comparison cannot succeed.
948 APInt Comp = CmpRHSV << ShAmtVal;
949 ConstantInt *ShiftedCmpRHS = ConstantInt::get(ICI.getContext(), Comp);
950 if (Shr->getOpcode() == Instruction::LShr)
951 Comp = Comp.lshr(ShAmtVal);
953 Comp = Comp.ashr(ShAmtVal);
955 if (Comp != CmpRHSV) { // Comparing against a bit that we know is zero.
956 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
957 Constant *Cst = ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
959 return ReplaceInstUsesWith(ICI, Cst);
962 // Otherwise, check to see if the bits shifted out are known to be zero.
963 // If so, we can compare against the unshifted value:
964 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
965 if (Shr->hasOneUse() && cast<BinaryOperator>(Shr)->isExact())
966 return new ICmpInst(ICI.getPredicate(), Shr->getOperand(0), ShiftedCmpRHS);
968 if (Shr->hasOneUse()) {
969 // Otherwise strength reduce the shift into an and.
970 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
971 Constant *Mask = ConstantInt::get(ICI.getContext(), Val);
973 Value *And = Builder->CreateAnd(Shr->getOperand(0),
974 Mask, Shr->getName()+".mask");
975 return new ICmpInst(ICI.getPredicate(), And, ShiftedCmpRHS);
981 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
983 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
986 const APInt &RHSV = RHS->getValue();
988 switch (LHSI->getOpcode()) {
989 case Instruction::Trunc:
990 if (ICI.isEquality() && LHSI->hasOneUse()) {
991 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
992 // of the high bits truncated out of x are known.
993 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
994 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
995 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
996 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
997 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
999 // If all the high bits are known, we can do this xform.
1000 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
1001 // Pull in the high bits from known-ones set.
1002 APInt NewRHS = RHS->getValue().zext(SrcBits);
1004 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1005 ConstantInt::get(ICI.getContext(), NewRHS));
1010 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
1011 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
1012 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
1014 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
1015 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
1016 Value *CompareVal = LHSI->getOperand(0);
1018 // If the sign bit of the XorCST is not set, there is no change to
1019 // the operation, just stop using the Xor.
1020 if (!XorCST->getValue().isNegative()) {
1021 ICI.setOperand(0, CompareVal);
1026 // Was the old condition true if the operand is positive?
1027 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
1029 // If so, the new one isn't.
1030 isTrueIfPositive ^= true;
1032 if (isTrueIfPositive)
1033 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
1036 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
1040 if (LHSI->hasOneUse()) {
1041 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
1042 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
1043 const APInt &SignBit = XorCST->getValue();
1044 ICmpInst::Predicate Pred = ICI.isSigned()
1045 ? ICI.getUnsignedPredicate()
1046 : ICI.getSignedPredicate();
1047 return new ICmpInst(Pred, LHSI->getOperand(0),
1048 ConstantInt::get(ICI.getContext(),
1052 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
1053 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
1054 const APInt &NotSignBit = XorCST->getValue();
1055 ICmpInst::Predicate Pred = ICI.isSigned()
1056 ? ICI.getUnsignedPredicate()
1057 : ICI.getSignedPredicate();
1058 Pred = ICI.getSwappedPredicate(Pred);
1059 return new ICmpInst(Pred, LHSI->getOperand(0),
1060 ConstantInt::get(ICI.getContext(),
1061 RHSV ^ NotSignBit));
1066 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
1067 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
1068 LHSI->getOperand(0)->hasOneUse()) {
1069 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
1071 // If the LHS is an AND of a truncating cast, we can widen the
1072 // and/compare to be the input width without changing the value
1073 // produced, eliminating a cast.
1074 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
1075 // We can do this transformation if either the AND constant does not
1076 // have its sign bit set or if it is an equality comparison.
1077 // Extending a relational comparison when we're checking the sign
1078 // bit would not work.
1079 if (Cast->hasOneUse() &&
1080 (ICI.isEquality() ||
1081 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
1083 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
1084 APInt NewCST = AndCST->getValue().zext(BitWidth);
1085 APInt NewCI = RHSV.zext(BitWidth);
1087 Builder->CreateAnd(Cast->getOperand(0),
1088 ConstantInt::get(ICI.getContext(), NewCST),
1090 return new ICmpInst(ICI.getPredicate(), NewAnd,
1091 ConstantInt::get(ICI.getContext(), NewCI));
1095 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
1096 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
1097 // happens a LOT in code produced by the C front-end, for bitfield
1099 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
1100 if (Shift && !Shift->isShift())
1104 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
1105 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
1106 const Type *AndTy = AndCST->getType(); // Type of the and.
1108 // We can fold this as long as we can't shift unknown bits
1109 // into the mask. This can only happen with signed shift
1110 // rights, as they sign-extend.
1112 bool CanFold = Shift->isLogicalShift();
1114 // To test for the bad case of the signed shr, see if any
1115 // of the bits shifted in could be tested after the mask.
1116 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
1117 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
1119 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
1120 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
1121 AndCST->getValue()) == 0)
1127 if (Shift->getOpcode() == Instruction::Shl)
1128 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
1130 NewCst = ConstantExpr::getShl(RHS, ShAmt);
1132 // Check to see if we are shifting out any of the bits being
1134 if (ConstantExpr::get(Shift->getOpcode(),
1135 NewCst, ShAmt) != RHS) {
1136 // If we shifted bits out, the fold is not going to work out.
1137 // As a special case, check to see if this means that the
1138 // result is always true or false now.
1139 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1140 return ReplaceInstUsesWith(ICI,
1141 ConstantInt::getFalse(ICI.getContext()));
1142 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
1143 return ReplaceInstUsesWith(ICI,
1144 ConstantInt::getTrue(ICI.getContext()));
1146 ICI.setOperand(1, NewCst);
1147 Constant *NewAndCST;
1148 if (Shift->getOpcode() == Instruction::Shl)
1149 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
1151 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
1152 LHSI->setOperand(1, NewAndCST);
1153 LHSI->setOperand(0, Shift->getOperand(0));
1154 Worklist.Add(Shift); // Shift is dead.
1160 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
1161 // preferable because it allows the C<<Y expression to be hoisted out
1162 // of a loop if Y is invariant and X is not.
1163 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
1164 ICI.isEquality() && !Shift->isArithmeticShift() &&
1165 !isa<Constant>(Shift->getOperand(0))) {
1168 if (Shift->getOpcode() == Instruction::LShr) {
1169 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
1171 // Insert a logical shift.
1172 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
1175 // Compute X & (C << Y).
1177 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
1179 ICI.setOperand(0, NewAnd);
1184 // Try to optimize things like "A[i]&42 == 0" to index computations.
1185 if (LoadInst *LI = dyn_cast<LoadInst>(LHSI->getOperand(0))) {
1186 if (GetElementPtrInst *GEP =
1187 dyn_cast<GetElementPtrInst>(LI->getOperand(0)))
1188 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
1189 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
1190 !LI->isVolatile() && isa<ConstantInt>(LHSI->getOperand(1))) {
1191 ConstantInt *C = cast<ConstantInt>(LHSI->getOperand(1));
1192 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV,ICI, C))
1198 case Instruction::Or: {
1199 if (!ICI.isEquality() || !RHS->isNullValue() || !LHSI->hasOneUse())
1202 if (match(LHSI, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) {
1203 // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0
1204 // -> and (icmp eq P, null), (icmp eq Q, null).
1205 Value *ICIP = Builder->CreateICmp(ICI.getPredicate(), P,
1206 Constant::getNullValue(P->getType()));
1207 Value *ICIQ = Builder->CreateICmp(ICI.getPredicate(), Q,
1208 Constant::getNullValue(Q->getType()));
1210 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
1211 Op = BinaryOperator::CreateAnd(ICIP, ICIQ);
1213 Op = BinaryOperator::CreateOr(ICIP, ICIQ);
1219 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
1220 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1223 uint32_t TypeBits = RHSV.getBitWidth();
1225 // Check that the shift amount is in range. If not, don't perform
1226 // undefined shifts. When the shift is visited it will be
1228 if (ShAmt->uge(TypeBits))
1231 if (ICI.isEquality()) {
1232 // If we are comparing against bits always shifted out, the
1233 // comparison cannot succeed.
1235 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
1237 if (Comp != RHS) {// Comparing against a bit that we know is zero.
1238 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1240 ConstantInt::get(Type::getInt1Ty(ICI.getContext()), IsICMP_NE);
1241 return ReplaceInstUsesWith(ICI, Cst);
1244 // If the shift is NUW, then it is just shifting out zeros, no need for an
1246 if (cast<BinaryOperator>(LHSI)->hasNoUnsignedWrap())
1247 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
1248 ConstantExpr::getLShr(RHS, ShAmt));
1250 if (LHSI->hasOneUse()) {
1251 // Otherwise strength reduce the shift into an and.
1252 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
1254 ConstantInt::get(ICI.getContext(), APInt::getLowBitsSet(TypeBits,
1255 TypeBits-ShAmtVal));
1258 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
1259 return new ICmpInst(ICI.getPredicate(), And,
1260 ConstantExpr::getLShr(RHS, ShAmt));
1264 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
1265 bool TrueIfSigned = false;
1266 if (LHSI->hasOneUse() &&
1267 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
1268 // (X << 31) <s 0 --> (X&1) != 0
1269 Constant *Mask = ConstantInt::get(ICI.getContext(), APInt(TypeBits, 1) <<
1270 (TypeBits-ShAmt->getZExtValue()-1));
1272 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
1273 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
1274 And, Constant::getNullValue(And->getType()));
1279 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
1280 case Instruction::AShr:
1281 // Only handle equality comparisons of shift-by-constant.
1282 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1283 if (Instruction *Res = FoldICmpShrCst(ICI, cast<BinaryOperator>(LHSI),
1288 case Instruction::SDiv:
1289 case Instruction::UDiv:
1290 // Fold: icmp pred ([us]div X, C1), C2 -> range test
1291 // Fold this div into the comparison, producing a range check.
1292 // Determine, based on the divide type, what the range is being
1293 // checked. If there is an overflow on the low or high side, remember
1294 // it, otherwise compute the range [low, hi) bounding the new value.
1295 // See: InsertRangeTest above for the kinds of replacements possible.
1296 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
1297 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
1302 case Instruction::Add:
1303 // Fold: icmp pred (add X, C1), C2
1304 if (!ICI.isEquality()) {
1305 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
1307 const APInt &LHSV = LHSC->getValue();
1309 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
1312 if (ICI.isSigned()) {
1313 if (CR.getLower().isSignBit()) {
1314 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
1315 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1316 } else if (CR.getUpper().isSignBit()) {
1317 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
1318 ConstantInt::get(ICI.getContext(),CR.getLower()));
1321 if (CR.getLower().isMinValue()) {
1322 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
1323 ConstantInt::get(ICI.getContext(),CR.getUpper()));
1324 } else if (CR.getUpper().isMinValue()) {
1325 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
1326 ConstantInt::get(ICI.getContext(),CR.getLower()));
1333 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
1334 if (ICI.isEquality()) {
1335 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
1337 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
1338 // the second operand is a constant, simplify a bit.
1339 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
1340 switch (BO->getOpcode()) {
1341 case Instruction::SRem:
1342 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
1343 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
1344 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
1345 if (V.sgt(1) && V.isPowerOf2()) {
1347 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
1349 return new ICmpInst(ICI.getPredicate(), NewRem,
1350 Constant::getNullValue(BO->getType()));
1354 case Instruction::Add:
1355 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
1356 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1357 if (BO->hasOneUse())
1358 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1359 ConstantExpr::getSub(RHS, BOp1C));
1360 } else if (RHSV == 0) {
1361 // Replace ((add A, B) != 0) with (A != -B) if A or B is
1362 // efficiently invertible, or if the add has just this one use.
1363 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
1365 if (Value *NegVal = dyn_castNegVal(BOp1))
1366 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
1367 else if (Value *NegVal = dyn_castNegVal(BOp0))
1368 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
1369 else if (BO->hasOneUse()) {
1370 Value *Neg = Builder->CreateNeg(BOp1);
1372 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
1376 case Instruction::Xor:
1377 // For the xor case, we can xor two constants together, eliminating
1378 // the explicit xor.
1379 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
1380 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1381 ConstantExpr::getXor(RHS, BOC));
1384 case Instruction::Sub:
1385 // Replace (([sub|xor] A, B) != 0) with (A != B)
1387 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
1391 case Instruction::Or:
1392 // If bits are being or'd in that are not present in the constant we
1393 // are comparing against, then the comparison could never succeed!
1394 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1395 Constant *NotCI = ConstantExpr::getNot(RHS);
1396 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
1397 return ReplaceInstUsesWith(ICI,
1398 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1403 case Instruction::And:
1404 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
1405 // If bits are being compared against that are and'd out, then the
1406 // comparison can never succeed!
1407 if ((RHSV & ~BOC->getValue()) != 0)
1408 return ReplaceInstUsesWith(ICI,
1409 ConstantInt::get(Type::getInt1Ty(ICI.getContext()),
1412 // If we have ((X & C) == C), turn it into ((X & C) != 0).
1413 if (RHS == BOC && RHSV.isPowerOf2())
1414 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
1415 ICmpInst::ICMP_NE, LHSI,
1416 Constant::getNullValue(RHS->getType()));
1418 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
1419 if (BOC->getValue().isSignBit()) {
1420 Value *X = BO->getOperand(0);
1421 Constant *Zero = Constant::getNullValue(X->getType());
1422 ICmpInst::Predicate pred = isICMP_NE ?
1423 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
1424 return new ICmpInst(pred, X, Zero);
1427 // ((X & ~7) == 0) --> X < 8
1428 if (RHSV == 0 && isHighOnes(BOC)) {
1429 Value *X = BO->getOperand(0);
1430 Constant *NegX = ConstantExpr::getNeg(BOC);
1431 ICmpInst::Predicate pred = isICMP_NE ?
1432 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
1433 return new ICmpInst(pred, X, NegX);
1438 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
1439 // Handle icmp {eq|ne} <intrinsic>, intcst.
1440 switch (II->getIntrinsicID()) {
1441 case Intrinsic::bswap:
1443 ICI.setOperand(0, II->getArgOperand(0));
1444 ICI.setOperand(1, ConstantInt::get(II->getContext(), RHSV.byteSwap()));
1446 case Intrinsic::ctlz:
1447 case Intrinsic::cttz:
1448 // ctz(A) == bitwidth(a) -> A == 0 and likewise for !=
1449 if (RHSV == RHS->getType()->getBitWidth()) {
1451 ICI.setOperand(0, II->getArgOperand(0));
1452 ICI.setOperand(1, ConstantInt::get(RHS->getType(), 0));
1456 case Intrinsic::ctpop:
1457 // popcount(A) == 0 -> A == 0 and likewise for !=
1458 if (RHS->isZero()) {
1460 ICI.setOperand(0, II->getArgOperand(0));
1461 ICI.setOperand(1, RHS);
1473 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
1474 /// We only handle extending casts so far.
1476 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
1477 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
1478 Value *LHSCIOp = LHSCI->getOperand(0);
1479 const Type *SrcTy = LHSCIOp->getType();
1480 const Type *DestTy = LHSCI->getType();
1483 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
1484 // integer type is the same size as the pointer type.
1485 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
1486 TD->getPointerSizeInBits() ==
1487 cast<IntegerType>(DestTy)->getBitWidth()) {
1489 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
1490 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
1491 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
1492 RHSOp = RHSC->getOperand(0);
1493 // If the pointer types don't match, insert a bitcast.
1494 if (LHSCIOp->getType() != RHSOp->getType())
1495 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
1499 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
1502 // The code below only handles extension cast instructions, so far.
1504 if (LHSCI->getOpcode() != Instruction::ZExt &&
1505 LHSCI->getOpcode() != Instruction::SExt)
1508 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
1509 bool isSignedCmp = ICI.isSigned();
1511 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
1512 // Not an extension from the same type?
1513 RHSCIOp = CI->getOperand(0);
1514 if (RHSCIOp->getType() != LHSCIOp->getType())
1517 // If the signedness of the two casts doesn't agree (i.e. one is a sext
1518 // and the other is a zext), then we can't handle this.
1519 if (CI->getOpcode() != LHSCI->getOpcode())
1522 // Deal with equality cases early.
1523 if (ICI.isEquality())
1524 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1526 // A signed comparison of sign extended values simplifies into a
1527 // signed comparison.
1528 if (isSignedCmp && isSignedExt)
1529 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
1531 // The other three cases all fold into an unsigned comparison.
1532 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
1535 // If we aren't dealing with a constant on the RHS, exit early
1536 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
1540 // Compute the constant that would happen if we truncated to SrcTy then
1541 // reextended to DestTy.
1542 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
1543 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
1546 // If the re-extended constant didn't change...
1548 // Deal with equality cases early.
1549 if (ICI.isEquality())
1550 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1552 // A signed comparison of sign extended values simplifies into a
1553 // signed comparison.
1554 if (isSignedExt && isSignedCmp)
1555 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
1557 // The other three cases all fold into an unsigned comparison.
1558 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, Res1);
1561 // The re-extended constant changed so the constant cannot be represented
1562 // in the shorter type. Consequently, we cannot emit a simple comparison.
1563 // All the cases that fold to true or false will have already been handled
1564 // by SimplifyICmpInst, so only deal with the tricky case.
1566 if (isSignedCmp || !isSignedExt)
1569 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
1570 // should have been folded away previously and not enter in here.
1572 // We're performing an unsigned comp with a sign extended value.
1573 // This is true if the input is >= 0. [aka >s -1]
1574 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
1575 Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
1577 // Finally, return the value computed.
1578 if (ICI.getPredicate() == ICmpInst::ICMP_ULT)
1579 return ReplaceInstUsesWith(ICI, Result);
1581 assert(ICI.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!");
1582 return BinaryOperator::CreateNot(Result);
1585 /// ProcessUGT_ADDCST_ADD - The caller has matched a pattern of the form:
1586 /// I = icmp ugt (add (add A, B), CI2), CI1
1587 /// If this is of the form:
1589 /// if (sum+128 >u 255)
1590 /// Then replace it with llvm.sadd.with.overflow.i8.
1592 static Instruction *ProcessUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
1593 ConstantInt *CI2, ConstantInt *CI1,
1595 // The transformation we're trying to do here is to transform this into an
1596 // llvm.sadd.with.overflow. To do this, we have to replace the original add
1597 // with a narrower add, and discard the add-with-constant that is part of the
1598 // range check (if we can't eliminate it, this isn't profitable).
1600 // In order to eliminate the add-with-constant, the compare can be its only
1602 Instruction *AddWithCst = cast<Instruction>(I.getOperand(0));
1603 if (!AddWithCst->hasOneUse()) return 0;
1605 // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow.
1606 if (!CI2->getValue().isPowerOf2()) return 0;
1607 unsigned NewWidth = CI2->getValue().countTrailingZeros();
1608 if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) return 0;
1610 // The width of the new add formed is 1 more than the bias.
1613 // Check to see that CI1 is an all-ones value with NewWidth bits.
1614 if (CI1->getBitWidth() == NewWidth ||
1615 CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth))
1618 // In order to replace the original add with a narrower
1619 // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant
1620 // and truncates that discard the high bits of the add. Verify that this is
1622 Instruction *OrigAdd = cast<Instruction>(AddWithCst->getOperand(0));
1623 for (Value::use_iterator UI = OrigAdd->use_begin(), E = OrigAdd->use_end();
1625 if (*UI == AddWithCst) continue;
1627 // Only accept truncates for now. We would really like a nice recursive
1628 // predicate like SimplifyDemandedBits, but which goes downwards the use-def
1629 // chain to see which bits of a value are actually demanded. If the
1630 // original add had another add which was then immediately truncated, we
1631 // could still do the transformation.
1632 TruncInst *TI = dyn_cast<TruncInst>(*UI);
1634 TI->getType()->getPrimitiveSizeInBits() > NewWidth) return 0;
1637 // If the pattern matches, truncate the inputs to the narrower type and
1638 // use the sadd_with_overflow intrinsic to efficiently compute both the
1639 // result and the overflow bit.
1640 Module *M = I.getParent()->getParent()->getParent();
1642 const Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth);
1643 Value *F = Intrinsic::getDeclaration(M, Intrinsic::sadd_with_overflow,
1646 InstCombiner::BuilderTy *Builder = IC.Builder;
1648 // Put the new code above the original add, in case there are any uses of the
1649 // add between the add and the compare.
1650 Builder->SetInsertPoint(OrigAdd);
1652 Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName()+".trunc");
1653 Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName()+".trunc");
1654 CallInst *Call = Builder->CreateCall2(F, TruncA, TruncB, "sadd");
1655 Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result");
1656 Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType());
1658 // The inner add was the result of the narrow add, zero extended to the
1659 // wider type. Replace it with the result computed by the intrinsic.
1660 IC.ReplaceInstUsesWith(*OrigAdd, ZExt);
1662 // The original icmp gets replaced with the overflow value.
1663 return ExtractValueInst::Create(Call, 1, "sadd.overflow");
1666 static Instruction *ProcessUAddIdiom(Instruction &I, Value *OrigAddV,
1668 // Don't bother doing this transformation for pointers, don't do it for
1670 if (!isa<IntegerType>(OrigAddV->getType())) return 0;
1672 // If the add is a constant expr, then we don't bother transforming it.
1673 Instruction *OrigAdd = dyn_cast<Instruction>(OrigAddV);
1674 if (OrigAdd == 0) return 0;
1676 Value *LHS = OrigAdd->getOperand(0), *RHS = OrigAdd->getOperand(1);
1678 // Put the new code above the original add, in case there are any uses of the
1679 // add between the add and the compare.
1680 InstCombiner::BuilderTy *Builder = IC.Builder;
1681 Builder->SetInsertPoint(OrigAdd);
1683 Module *M = I.getParent()->getParent()->getParent();
1684 const Type *Ty = LHS->getType();
1685 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, &Ty,1);
1686 CallInst *Call = Builder->CreateCall2(F, LHS, RHS, "uadd");
1687 Value *Add = Builder->CreateExtractValue(Call, 0);
1689 IC.ReplaceInstUsesWith(*OrigAdd, Add);
1691 // The original icmp gets replaced with the overflow value.
1692 return ExtractValueInst::Create(Call, 1, "uadd.overflow");
1695 // DemandedBitsLHSMask - When performing a comparison against a constant,
1696 // it is possible that not all the bits in the LHS are demanded. This helper
1697 // method computes the mask that IS demanded.
1698 static APInt DemandedBitsLHSMask(ICmpInst &I,
1699 unsigned BitWidth, bool isSignCheck) {
1701 return APInt::getSignBit(BitWidth);
1703 ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1));
1704 if (!CI) return APInt::getAllOnesValue(BitWidth);
1705 const APInt &RHS = CI->getValue();
1707 switch (I.getPredicate()) {
1708 // For a UGT comparison, we don't care about any bits that
1709 // correspond to the trailing ones of the comparand. The value of these
1710 // bits doesn't impact the outcome of the comparison, because any value
1711 // greater than the RHS must differ in a bit higher than these due to carry.
1712 case ICmpInst::ICMP_UGT: {
1713 unsigned trailingOnes = RHS.countTrailingOnes();
1714 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes);
1718 // Similarly, for a ULT comparison, we don't care about the trailing zeros.
1719 // Any value less than the RHS must differ in a higher bit because of carries.
1720 case ICmpInst::ICMP_ULT: {
1721 unsigned trailingZeros = RHS.countTrailingZeros();
1722 APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros);
1727 return APInt::getAllOnesValue(BitWidth);
1732 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
1733 bool Changed = false;
1734 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1736 /// Orders the operands of the compare so that they are listed from most
1737 /// complex to least complex. This puts constants before unary operators,
1738 /// before binary operators.
1739 if (getComplexity(Op0) < getComplexity(Op1)) {
1741 std::swap(Op0, Op1);
1745 if (Value *V = SimplifyICmpInst(I.getPredicate(), Op0, Op1, TD))
1746 return ReplaceInstUsesWith(I, V);
1748 const Type *Ty = Op0->getType();
1750 // icmp's with boolean values can always be turned into bitwise operations
1751 if (Ty->isIntegerTy(1)) {
1752 switch (I.getPredicate()) {
1753 default: llvm_unreachable("Invalid icmp instruction!");
1754 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
1755 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
1756 return BinaryOperator::CreateNot(Xor);
1758 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
1759 return BinaryOperator::CreateXor(Op0, Op1);
1761 case ICmpInst::ICMP_UGT:
1762 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
1764 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
1765 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1766 return BinaryOperator::CreateAnd(Not, Op1);
1768 case ICmpInst::ICMP_SGT:
1769 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
1771 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
1772 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1773 return BinaryOperator::CreateAnd(Not, Op0);
1775 case ICmpInst::ICMP_UGE:
1776 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
1778 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
1779 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
1780 return BinaryOperator::CreateOr(Not, Op1);
1782 case ICmpInst::ICMP_SGE:
1783 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
1785 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
1786 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
1787 return BinaryOperator::CreateOr(Not, Op0);
1792 unsigned BitWidth = 0;
1793 if (Ty->isIntOrIntVectorTy())
1794 BitWidth = Ty->getScalarSizeInBits();
1795 else if (TD) // Pointers require TD info to get their size.
1796 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
1798 bool isSignBit = false;
1800 // See if we are doing a comparison with a constant.
1801 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1802 Value *A = 0, *B = 0;
1804 // Match the following pattern, which is a common idiom when writing
1805 // overflow-safe integer arithmetic function. The source performs an
1806 // addition in wider type, and explicitly checks for overflow using
1807 // comparisons against INT_MIN and INT_MAX. Simplify this by using the
1808 // sadd_with_overflow intrinsic.
1810 // TODO: This could probably be generalized to handle other overflow-safe
1811 // operations if we worked out the formulas to compute the appropriate
1815 // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8
1817 ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
1818 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
1819 match(Op0, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
1820 if (Instruction *Res = ProcessUGT_ADDCST_ADD(I, A, B, CI2, CI, *this))
1824 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
1825 if (I.isEquality() && CI->isZero() &&
1826 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
1827 // (icmp cond A B) if cond is equality
1828 return new ICmpInst(I.getPredicate(), A, B);
1831 // If we have an icmp le or icmp ge instruction, turn it into the
1832 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
1833 // them being folded in the code below. The SimplifyICmpInst code has
1834 // already handled the edge cases for us, so we just assert on them.
1835 switch (I.getPredicate()) {
1837 case ICmpInst::ICMP_ULE:
1838 assert(!CI->isMaxValue(false)); // A <=u MAX -> TRUE
1839 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
1840 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1841 case ICmpInst::ICMP_SLE:
1842 assert(!CI->isMaxValue(true)); // A <=s MAX -> TRUE
1843 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
1844 ConstantInt::get(CI->getContext(), CI->getValue()+1));
1845 case ICmpInst::ICMP_UGE:
1846 assert(!CI->isMinValue(false)); // A >=u MIN -> TRUE
1847 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
1848 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1849 case ICmpInst::ICMP_SGE:
1850 assert(!CI->isMinValue(true)); // A >=s MIN -> TRUE
1851 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1852 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1855 // If this comparison is a normal comparison, it demands all
1856 // bits, if it is a sign bit comparison, it only demands the sign bit.
1858 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
1861 // See if we can fold the comparison based on range information we can get
1862 // by checking whether bits are known to be zero or one in the input.
1863 if (BitWidth != 0) {
1864 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
1865 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
1867 if (SimplifyDemandedBits(I.getOperandUse(0),
1868 DemandedBitsLHSMask(I, BitWidth, isSignBit),
1869 Op0KnownZero, Op0KnownOne, 0))
1871 if (SimplifyDemandedBits(I.getOperandUse(1),
1872 APInt::getAllOnesValue(BitWidth),
1873 Op1KnownZero, Op1KnownOne, 0))
1876 // Given the known and unknown bits, compute a range that the LHS could be
1877 // in. Compute the Min, Max and RHS values based on the known bits. For the
1878 // EQ and NE we use unsigned values.
1879 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
1880 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
1882 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1884 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1887 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
1889 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
1893 // If Min and Max are known to be the same, then SimplifyDemandedBits
1894 // figured out that the LHS is a constant. Just constant fold this now so
1895 // that code below can assume that Min != Max.
1896 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
1897 return new ICmpInst(I.getPredicate(),
1898 ConstantInt::get(I.getContext(), Op0Min), Op1);
1899 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
1900 return new ICmpInst(I.getPredicate(), Op0,
1901 ConstantInt::get(I.getContext(), Op1Min));
1903 // Based on the range information we know about the LHS, see if we can
1904 // simplify this comparison. For example, (x&4) < 8 is always true.
1905 switch (I.getPredicate()) {
1906 default: llvm_unreachable("Unknown icmp opcode!");
1907 case ICmpInst::ICMP_EQ: {
1908 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1909 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1911 // If all bits are known zero except for one, then we know at most one
1912 // bit is set. If the comparison is against zero, then this is a check
1913 // to see if *that* bit is set.
1914 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1915 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1916 // If the LHS is an AND with the same constant, look through it.
1918 ConstantInt *LHSC = 0;
1919 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1920 LHSC->getValue() != Op0KnownZeroInverted)
1923 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1924 // then turn "((1 << x)&8) == 0" into "x != 3".
1926 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1927 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1928 return new ICmpInst(ICmpInst::ICMP_NE, X,
1929 ConstantInt::get(X->getType(), CmpVal));
1932 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1933 // then turn "((8 >>u x)&1) == 0" into "x != 3".
1935 if (Op0KnownZeroInverted == 1 &&
1936 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1937 return new ICmpInst(ICmpInst::ICMP_NE, X,
1938 ConstantInt::get(X->getType(),
1939 CI->countTrailingZeros()));
1944 case ICmpInst::ICMP_NE: {
1945 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
1946 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1948 // If all bits are known zero except for one, then we know at most one
1949 // bit is set. If the comparison is against zero, then this is a check
1950 // to see if *that* bit is set.
1951 APInt Op0KnownZeroInverted = ~Op0KnownZero;
1952 if (~Op1KnownZero == 0 && Op0KnownZeroInverted.isPowerOf2()) {
1953 // If the LHS is an AND with the same constant, look through it.
1955 ConstantInt *LHSC = 0;
1956 if (!match(Op0, m_And(m_Value(LHS), m_ConstantInt(LHSC))) ||
1957 LHSC->getValue() != Op0KnownZeroInverted)
1960 // If the LHS is 1 << x, and we know the result is a power of 2 like 8,
1961 // then turn "((1 << x)&8) != 0" into "x == 3".
1963 if (match(LHS, m_Shl(m_One(), m_Value(X)))) {
1964 unsigned CmpVal = Op0KnownZeroInverted.countTrailingZeros();
1965 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1966 ConstantInt::get(X->getType(), CmpVal));
1969 // If the LHS is 8 >>u x, and we know the result is a power of 2 like 1,
1970 // then turn "((8 >>u x)&1) != 0" into "x == 3".
1972 if (Op0KnownZeroInverted == 1 &&
1973 match(LHS, m_LShr(m_Power2(CI), m_Value(X))))
1974 return new ICmpInst(ICmpInst::ICMP_EQ, X,
1975 ConstantInt::get(X->getType(),
1976 CI->countTrailingZeros()));
1981 case ICmpInst::ICMP_ULT:
1982 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
1983 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
1984 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
1985 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
1986 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
1987 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
1988 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
1989 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
1990 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
1991 ConstantInt::get(CI->getContext(), CI->getValue()-1));
1993 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
1994 if (CI->isMinValue(true))
1995 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
1996 Constant::getAllOnesValue(Op0->getType()));
1999 case ICmpInst::ICMP_UGT:
2000 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
2001 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2002 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
2003 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2005 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
2006 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2007 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2008 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
2009 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2010 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2012 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
2013 if (CI->isMaxValue(true))
2014 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
2015 Constant::getNullValue(Op0->getType()));
2018 case ICmpInst::ICMP_SLT:
2019 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
2020 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2021 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
2022 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2023 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
2024 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2025 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2026 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
2027 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2028 ConstantInt::get(CI->getContext(), CI->getValue()-1));
2031 case ICmpInst::ICMP_SGT:
2032 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
2033 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2034 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
2035 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2037 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
2038 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
2039 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2040 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
2041 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
2042 ConstantInt::get(CI->getContext(), CI->getValue()+1));
2045 case ICmpInst::ICMP_SGE:
2046 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
2047 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
2048 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2049 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
2050 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2052 case ICmpInst::ICMP_SLE:
2053 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
2054 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
2055 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2056 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
2057 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2059 case ICmpInst::ICMP_UGE:
2060 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
2061 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
2062 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2063 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
2064 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2066 case ICmpInst::ICMP_ULE:
2067 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
2068 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
2069 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2070 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
2071 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2075 // Turn a signed comparison into an unsigned one if both operands
2076 // are known to have the same sign.
2078 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
2079 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
2080 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
2083 // Test if the ICmpInst instruction is used exclusively by a select as
2084 // part of a minimum or maximum operation. If so, refrain from doing
2085 // any other folding. This helps out other analyses which understand
2086 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
2087 // and CodeGen. And in this case, at least one of the comparison
2088 // operands has at least one user besides the compare (the select),
2089 // which would often largely negate the benefit of folding anyway.
2091 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
2092 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
2093 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
2096 // See if we are doing a comparison between a constant and an instruction that
2097 // can be folded into the comparison.
2098 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2099 // Since the RHS is a ConstantInt (CI), if the left hand side is an
2100 // instruction, see if that instruction also has constants so that the
2101 // instruction can be folded into the icmp
2102 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2103 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
2107 // Handle icmp with constant (but not simple integer constant) RHS
2108 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2109 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2110 switch (LHSI->getOpcode()) {
2111 case Instruction::GetElementPtr:
2112 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
2113 if (RHSC->isNullValue() &&
2114 cast<GetElementPtrInst>(LHSI)->hasAllZeroIndices())
2115 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2116 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2118 case Instruction::PHI:
2119 // Only fold icmp into the PHI if the phi and icmp are in the same
2120 // block. If in the same block, we're encouraging jump threading. If
2121 // not, we are just pessimizing the code by making an i1 phi.
2122 if (LHSI->getParent() == I.getParent())
2123 if (Instruction *NV = FoldOpIntoPhi(I))
2126 case Instruction::Select: {
2127 // If either operand of the select is a constant, we can fold the
2128 // comparison into the select arms, which will cause one to be
2129 // constant folded and the select turned into a bitwise or.
2130 Value *Op1 = 0, *Op2 = 0;
2131 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1)))
2132 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2133 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2)))
2134 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
2136 // We only want to perform this transformation if it will not lead to
2137 // additional code. This is true if either both sides of the select
2138 // fold to a constant (in which case the icmp is replaced with a select
2139 // which will usually simplify) or this is the only user of the
2140 // select (in which case we are trading a select+icmp for a simpler
2142 if ((Op1 && Op2) || (LHSI->hasOneUse() && (Op1 || Op2))) {
2144 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
2147 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
2149 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2153 case Instruction::IntToPtr:
2154 // icmp pred inttoptr(X), null -> icmp pred X, 0
2155 if (RHSC->isNullValue() && TD &&
2156 TD->getIntPtrType(RHSC->getContext()) ==
2157 LHSI->getOperand(0)->getType())
2158 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
2159 Constant::getNullValue(LHSI->getOperand(0)->getType()));
2162 case Instruction::Load:
2163 // Try to optimize things like "A[i] > 4" to index computations.
2164 if (GetElementPtrInst *GEP =
2165 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2166 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2167 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2168 !cast<LoadInst>(LHSI)->isVolatile())
2169 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2176 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
2177 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
2178 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
2180 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
2181 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
2182 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
2185 // Test to see if the operands of the icmp are casted versions of other
2186 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
2188 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
2189 if (Op0->getType()->isPointerTy() &&
2190 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
2191 // We keep moving the cast from the left operand over to the right
2192 // operand, where it can often be eliminated completely.
2193 Op0 = CI->getOperand(0);
2195 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
2196 // so eliminate it as well.
2197 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
2198 Op1 = CI2->getOperand(0);
2200 // If Op1 is a constant, we can fold the cast into the constant.
2201 if (Op0->getType() != Op1->getType()) {
2202 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
2203 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
2205 // Otherwise, cast the RHS right before the icmp
2206 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
2209 return new ICmpInst(I.getPredicate(), Op0, Op1);
2213 if (isa<CastInst>(Op0)) {
2214 // Handle the special case of: icmp (cast bool to X), <cst>
2215 // This comes up when you have code like
2218 // For generality, we handle any zero-extension of any operand comparison
2219 // with a constant or another cast from the same type.
2220 if (isa<Constant>(Op1) || isa<CastInst>(Op1))
2221 if (Instruction *R = visitICmpInstWithCastAndCast(I))
2225 // See if it's the same type of instruction on the left and right.
2226 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2227 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2228 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
2229 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
2230 switch (Op0I->getOpcode()) {
2232 case Instruction::Add:
2233 case Instruction::Sub:
2234 case Instruction::Xor:
2235 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
2236 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
2237 Op1I->getOperand(0));
2238 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
2239 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2240 if (CI->getValue().isSignBit()) {
2241 ICmpInst::Predicate Pred = I.isSigned()
2242 ? I.getUnsignedPredicate()
2243 : I.getSignedPredicate();
2244 return new ICmpInst(Pred, Op0I->getOperand(0),
2245 Op1I->getOperand(0));
2248 if (CI->getValue().isMaxSignedValue()) {
2249 ICmpInst::Predicate Pred = I.isSigned()
2250 ? I.getUnsignedPredicate()
2251 : I.getSignedPredicate();
2252 Pred = I.getSwappedPredicate(Pred);
2253 return new ICmpInst(Pred, Op0I->getOperand(0),
2254 Op1I->getOperand(0));
2258 case Instruction::Mul:
2259 if (!I.isEquality())
2262 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2263 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
2264 // Mask = -1 >> count-trailing-zeros(Cst).
2265 if (!CI->isZero() && !CI->isOne()) {
2266 const APInt &AP = CI->getValue();
2267 ConstantInt *Mask = ConstantInt::get(I.getContext(),
2268 APInt::getLowBitsSet(AP.getBitWidth(),
2270 AP.countTrailingZeros()));
2271 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
2272 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
2273 return new ICmpInst(I.getPredicate(), And1, And2);
2283 // ~x < ~y --> y < x
2284 // ~x < cst --> ~cst < x
2285 if (match(Op0, m_Not(m_Value(A)))) {
2286 if (match(Op1, m_Not(m_Value(B))))
2287 return new ICmpInst(I.getPredicate(), B, A);
2288 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1))
2289 return new ICmpInst(I.getPredicate(), ConstantExpr::getNot(RHSC), A);
2292 // (a+b) <u a --> llvm.uadd.with.overflow.
2293 // (a+b) <u b --> llvm.uadd.with.overflow.
2294 if (I.getPredicate() == ICmpInst::ICMP_ULT &&
2295 match(Op0, m_Add(m_Value(A), m_Value(B))) &&
2296 (Op1 == A || Op1 == B))
2297 if (Instruction *R = ProcessUAddIdiom(I, Op0, *this))
2300 // a >u (a+b) --> llvm.uadd.with.overflow.
2301 // b >u (a+b) --> llvm.uadd.with.overflow.
2302 if (I.getPredicate() == ICmpInst::ICMP_UGT &&
2303 match(Op1, m_Add(m_Value(A), m_Value(B))) &&
2304 (Op0 == A || Op0 == B))
2305 if (Instruction *R = ProcessUAddIdiom(I, Op1, *this))
2309 if (I.isEquality()) {
2310 Value *A, *B, *C, *D;
2312 // -x == -y --> x == y
2313 if (match(Op0, m_Neg(m_Value(A))) &&
2314 match(Op1, m_Neg(m_Value(B))))
2315 return new ICmpInst(I.getPredicate(), A, B);
2317 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
2318 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
2319 Value *OtherVal = A == Op1 ? B : A;
2320 return new ICmpInst(I.getPredicate(), OtherVal,
2321 Constant::getNullValue(A->getType()));
2324 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
2325 // A^c1 == C^c2 --> A == C^(c1^c2)
2326 ConstantInt *C1, *C2;
2327 if (match(B, m_ConstantInt(C1)) &&
2328 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
2329 Constant *NC = ConstantInt::get(I.getContext(),
2330 C1->getValue() ^ C2->getValue());
2331 Value *Xor = Builder->CreateXor(C, NC, "tmp");
2332 return new ICmpInst(I.getPredicate(), A, Xor);
2335 // A^B == A^D -> B == D
2336 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
2337 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
2338 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
2339 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
2343 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
2344 (A == Op0 || B == Op0)) {
2345 // A == (A^B) -> B == 0
2346 Value *OtherVal = A == Op0 ? B : A;
2347 return new ICmpInst(I.getPredicate(), OtherVal,
2348 Constant::getNullValue(A->getType()));
2351 // (A-B) == A -> B == 0
2352 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
2353 return new ICmpInst(I.getPredicate(), B,
2354 Constant::getNullValue(B->getType()));
2356 // A == (A-B) -> B == 0
2357 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
2358 return new ICmpInst(I.getPredicate(), B,
2359 Constant::getNullValue(B->getType()));
2361 // (A+B) == A -> B == 0
2362 if (match(Op0, m_Add(m_Specific(Op1), m_Value(B))) ||
2363 match(Op0, m_Add(m_Value(B), m_Specific(Op1))))
2364 return new ICmpInst(I.getPredicate(), B,
2365 Constant::getNullValue(B->getType()));
2367 // A == (A+B) -> B == 0
2368 if (match(Op1, m_Add(m_Specific(Op0), m_Value(B))) ||
2369 match(Op1, m_Add(m_Value(B), m_Specific(Op0))))
2370 return new ICmpInst(I.getPredicate(), B,
2371 Constant::getNullValue(B->getType()));
2373 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
2374 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2375 match(Op0, m_And(m_Value(A), m_Value(B))) &&
2376 match(Op1, m_And(m_Value(C), m_Value(D)))) {
2377 Value *X = 0, *Y = 0, *Z = 0;
2380 X = B; Y = D; Z = A;
2381 } else if (A == D) {
2382 X = B; Y = C; Z = A;
2383 } else if (B == C) {
2384 X = A; Y = D; Z = B;
2385 } else if (B == D) {
2386 X = A; Y = C; Z = B;
2389 if (X) { // Build (X^Y) & Z
2390 Op1 = Builder->CreateXor(X, Y, "tmp");
2391 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
2392 I.setOperand(0, Op1);
2393 I.setOperand(1, Constant::getNullValue(Op1->getType()));
2400 Value *X; ConstantInt *Cst;
2402 if (match(Op0, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op1 == X)
2403 return FoldICmpAddOpCst(I, X, Cst, I.getPredicate(), Op0);
2406 if (match(Op1, m_Add(m_Value(X), m_ConstantInt(Cst))) && Op0 == X)
2407 return FoldICmpAddOpCst(I, X, Cst, I.getSwappedPredicate(), Op1);
2409 return Changed ? &I : 0;
2417 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
2419 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
2422 if (!isa<ConstantFP>(RHSC)) return 0;
2423 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
2425 // Get the width of the mantissa. We don't want to hack on conversions that
2426 // might lose information from the integer, e.g. "i64 -> float"
2427 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
2428 if (MantissaWidth == -1) return 0; // Unknown.
2430 // Check to see that the input is converted from an integer type that is small
2431 // enough that preserves all bits. TODO: check here for "known" sign bits.
2432 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
2433 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
2435 // If this is a uitofp instruction, we need an extra bit to hold the sign.
2436 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
2440 // If the conversion would lose info, don't hack on this.
2441 if ((int)InputSize > MantissaWidth)
2444 // Otherwise, we can potentially simplify the comparison. We know that it
2445 // will always come through as an integer value and we know the constant is
2446 // not a NAN (it would have been previously simplified).
2447 assert(!RHS.isNaN() && "NaN comparison not already folded!");
2449 ICmpInst::Predicate Pred;
2450 switch (I.getPredicate()) {
2451 default: llvm_unreachable("Unexpected predicate!");
2452 case FCmpInst::FCMP_UEQ:
2453 case FCmpInst::FCMP_OEQ:
2454 Pred = ICmpInst::ICMP_EQ;
2456 case FCmpInst::FCMP_UGT:
2457 case FCmpInst::FCMP_OGT:
2458 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
2460 case FCmpInst::FCMP_UGE:
2461 case FCmpInst::FCMP_OGE:
2462 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
2464 case FCmpInst::FCMP_ULT:
2465 case FCmpInst::FCMP_OLT:
2466 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
2468 case FCmpInst::FCMP_ULE:
2469 case FCmpInst::FCMP_OLE:
2470 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
2472 case FCmpInst::FCMP_UNE:
2473 case FCmpInst::FCMP_ONE:
2474 Pred = ICmpInst::ICMP_NE;
2476 case FCmpInst::FCMP_ORD:
2477 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2478 case FCmpInst::FCMP_UNO:
2479 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2482 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
2484 // Now we know that the APFloat is a normal number, zero or inf.
2486 // See if the FP constant is too large for the integer. For example,
2487 // comparing an i8 to 300.0.
2488 unsigned IntWidth = IntTy->getScalarSizeInBits();
2491 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
2492 // and large values.
2493 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
2494 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
2495 APFloat::rmNearestTiesToEven);
2496 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
2497 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
2498 Pred == ICmpInst::ICMP_SLE)
2499 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2500 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2503 // If the RHS value is > UnsignedMax, fold the comparison. This handles
2504 // +INF and large values.
2505 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
2506 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
2507 APFloat::rmNearestTiesToEven);
2508 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
2509 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
2510 Pred == ICmpInst::ICMP_ULE)
2511 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2512 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2517 // See if the RHS value is < SignedMin.
2518 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
2519 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
2520 APFloat::rmNearestTiesToEven);
2521 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
2522 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
2523 Pred == ICmpInst::ICMP_SGE)
2524 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2525 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2529 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
2530 // [0, UMAX], but it may still be fractional. See if it is fractional by
2531 // casting the FP value to the integer value and back, checking for equality.
2532 // Don't do this for zero, because -0.0 is not fractional.
2533 Constant *RHSInt = LHSUnsigned
2534 ? ConstantExpr::getFPToUI(RHSC, IntTy)
2535 : ConstantExpr::getFPToSI(RHSC, IntTy);
2536 if (!RHS.isZero()) {
2537 bool Equal = LHSUnsigned
2538 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
2539 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
2541 // If we had a comparison against a fractional value, we have to adjust
2542 // the compare predicate and sometimes the value. RHSC is rounded towards
2543 // zero at this point.
2545 default: llvm_unreachable("Unexpected integer comparison!");
2546 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
2547 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2548 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
2549 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2550 case ICmpInst::ICMP_ULE:
2551 // (float)int <= 4.4 --> int <= 4
2552 // (float)int <= -4.4 --> false
2553 if (RHS.isNegative())
2554 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2556 case ICmpInst::ICMP_SLE:
2557 // (float)int <= 4.4 --> int <= 4
2558 // (float)int <= -4.4 --> int < -4
2559 if (RHS.isNegative())
2560 Pred = ICmpInst::ICMP_SLT;
2562 case ICmpInst::ICMP_ULT:
2563 // (float)int < -4.4 --> false
2564 // (float)int < 4.4 --> int <= 4
2565 if (RHS.isNegative())
2566 return ReplaceInstUsesWith(I, ConstantInt::getFalse(I.getContext()));
2567 Pred = ICmpInst::ICMP_ULE;
2569 case ICmpInst::ICMP_SLT:
2570 // (float)int < -4.4 --> int < -4
2571 // (float)int < 4.4 --> int <= 4
2572 if (!RHS.isNegative())
2573 Pred = ICmpInst::ICMP_SLE;
2575 case ICmpInst::ICMP_UGT:
2576 // (float)int > 4.4 --> int > 4
2577 // (float)int > -4.4 --> true
2578 if (RHS.isNegative())
2579 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2581 case ICmpInst::ICMP_SGT:
2582 // (float)int > 4.4 --> int > 4
2583 // (float)int > -4.4 --> int >= -4
2584 if (RHS.isNegative())
2585 Pred = ICmpInst::ICMP_SGE;
2587 case ICmpInst::ICMP_UGE:
2588 // (float)int >= -4.4 --> true
2589 // (float)int >= 4.4 --> int > 4
2590 if (!RHS.isNegative())
2591 return ReplaceInstUsesWith(I, ConstantInt::getTrue(I.getContext()));
2592 Pred = ICmpInst::ICMP_UGT;
2594 case ICmpInst::ICMP_SGE:
2595 // (float)int >= -4.4 --> int >= -4
2596 // (float)int >= 4.4 --> int > 4
2597 if (!RHS.isNegative())
2598 Pred = ICmpInst::ICMP_SGT;
2604 // Lower this FP comparison into an appropriate integer version of the
2606 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
2609 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
2610 bool Changed = false;
2612 /// Orders the operands of the compare so that they are listed from most
2613 /// complex to least complex. This puts constants before unary operators,
2614 /// before binary operators.
2615 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1))) {
2620 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2622 if (Value *V = SimplifyFCmpInst(I.getPredicate(), Op0, Op1, TD))
2623 return ReplaceInstUsesWith(I, V);
2625 // Simplify 'fcmp pred X, X'
2627 switch (I.getPredicate()) {
2628 default: llvm_unreachable("Unknown predicate!");
2629 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
2630 case FCmpInst::FCMP_ULT: // True if unordered or less than
2631 case FCmpInst::FCMP_UGT: // True if unordered or greater than
2632 case FCmpInst::FCMP_UNE: // True if unordered or not equal
2633 // Canonicalize these to be 'fcmp uno %X, 0.0'.
2634 I.setPredicate(FCmpInst::FCMP_UNO);
2635 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2638 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
2639 case FCmpInst::FCMP_OEQ: // True if ordered and equal
2640 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
2641 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
2642 // Canonicalize these to be 'fcmp ord %X, 0.0'.
2643 I.setPredicate(FCmpInst::FCMP_ORD);
2644 I.setOperand(1, Constant::getNullValue(Op0->getType()));
2649 // Handle fcmp with constant RHS
2650 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
2651 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
2652 switch (LHSI->getOpcode()) {
2653 case Instruction::PHI:
2654 // Only fold fcmp into the PHI if the phi and fcmp are in the same
2655 // block. If in the same block, we're encouraging jump threading. If
2656 // not, we are just pessimizing the code by making an i1 phi.
2657 if (LHSI->getParent() == I.getParent())
2658 if (Instruction *NV = FoldOpIntoPhi(I))
2661 case Instruction::SIToFP:
2662 case Instruction::UIToFP:
2663 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
2666 case Instruction::Select: {
2667 // If either operand of the select is a constant, we can fold the
2668 // comparison into the select arms, which will cause one to be
2669 // constant folded and the select turned into a bitwise or.
2670 Value *Op1 = 0, *Op2 = 0;
2671 if (LHSI->hasOneUse()) {
2672 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
2673 // Fold the known value into the constant operand.
2674 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2675 // Insert a new FCmp of the other select operand.
2676 Op2 = Builder->CreateFCmp(I.getPredicate(),
2677 LHSI->getOperand(2), RHSC, I.getName());
2678 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
2679 // Fold the known value into the constant operand.
2680 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
2681 // Insert a new FCmp of the other select operand.
2682 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
2688 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
2691 case Instruction::Load:
2692 if (GetElementPtrInst *GEP =
2693 dyn_cast<GetElementPtrInst>(LHSI->getOperand(0))) {
2694 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)))
2695 if (GV->isConstant() && GV->hasDefinitiveInitializer() &&
2696 !cast<LoadInst>(LHSI)->isVolatile())
2697 if (Instruction *Res = FoldCmpLoadFromIndexedGlobal(GEP, GV, I))
2704 return Changed ? &I : 0;